•  
  •  
 

Corresponding Author

Heba M. El-Hoseny

Document Type

Article

Keywords

Crowd counting, Hybrid attention, Mechanism, Curriculum learning, Context-aware network (CANNet)

Abstract

Accurate crowd counting remains a major challenge in computer vision due to perspective distortion, background clutter, occlusion, and poor illumination, which degrade feature representation and limit model generalization. This study introduces a robust and well-structured approach that integrates hybrid attention mechanisms with curriculum learning to enhance accuracy and stability in complex crowd scenarios. Gamma correction and CLAHE preprocessing enhance image visibility, facilitating the reliable detection of small and distant individuals under challenging lighting conditions. A mixed attention mechanism further enhances feature discrimination by emphasizing crowd-relevant regions while suppressing background noise. In addition, curriculum learning, guided by scoring and pacing functions, progressively trains the network from easy to difficult samples to ensure stable convergence and stronger generalization. Extensive experiments on four benchmark datasets, ShanghaiTech Part A/B, JHU-CROWD++, and UCF-QNRF, demonstrate the effectiveness of the proposed Smart CANNet. On Part A, the model achieved 42.56 Mean Absolute Error (MAE), 160.21 Mean Squared Error (MSE), 0.578 Structural Similarity Index Measure (SSIM), and 71.10 Peak Signal-to-Noise Ratio (PSNR). On Part B, it obtained 17.24 MAE, 80.36 MSE, 0.4935 SSIM, and 75.86 PSNR. On JHU-CROWD++, it reached 69.63 MAE and 290.3 MSE. Compared with MC-CNN and CSRNet, Smart CANNet reduced MAE by 19.7% and 11%, respectively, confirming its superior accuracy, robustness, and scalability across diverse real-world crowd scenes.

Download

Share

COinS